src/rdb.c

   1 #include "redis.h"
   2 #include "lzf.h"    /* LZF compression library */
   3
   4 #include <math.h>
   5 #include <sys/types.h>
   6 #include <sys/time.h>
   7 #include <sys/resource.h>
   8 #include <sys/wait.h>
   9 #include <arpa/inet.h>
  10 #include <sys/stat.h>
  11
  12 /* Convenience wrapper around fwrite, that returns the number of bytes written
  13  * to the file instead of the number of objects (see fwrite(3)) and -1 in the
  14  * case of an error. It also supports a NULL *fp to skip writing altogether
  15  * instead of writing to /dev/null. */
  16 static int rdbWriteRaw(FILE *fp, void *p, size_t len) {
  17     if (fp != NULL && fwrite(p,len,1,fp) == 0) return -1;
  18     return len;
  19 }
  20
  21 int rdbSaveType(FILE *fp, unsigned char type) {
  22     return rdbWriteRaw(fp,&type,1);
  23 }
  24
  25 int rdbSaveTime(FILE *fp, time_t t) {
  26     int32_t t32 = (int32_t) t;
  27     return rdbWriteRaw(fp,&t32,4);
  28 }
  29
  30 /* check rdbLoadLen() comments for more info */
  31 int rdbSaveLen(FILE *fp, uint32_t len) {
  32     unsigned char buf[2];
  33     int nwritten;
  34
  35     if (len < (1<<6)) {
  36         /* Save a 6 bit len */
  37         buf[0] = (len&0xFF)|(REDIS_RDB_6BITLEN<<6);
  38         if (rdbWriteRaw(fp,buf,1) == -1) return -1;
  39         nwritten = 1;
  40     } else if (len < (1<<14)) {
  41         /* Save a 14 bit len */
  42         buf[0] = ((len>>8)&0xFF)|(REDIS_RDB_14BITLEN<<6);
  43         buf[1] = len&0xFF;
  44         if (rdbWriteRaw(fp,buf,2) == -1) return -1;
  45         nwritten = 2;
  46     } else {
  47         /* Save a 32 bit len */
  48         buf[0] = (REDIS_RDB_32BITLEN<<6);
  49         if (rdbWriteRaw(fp,buf,1) == -1) return -1;
  50         len = htonl(len);
  51         if (rdbWriteRaw(fp,&len,4) == -1) return -1;
  52         nwritten = 1+4;
  53     }
  54     return nwritten;
  55 }
  56
  57 /* Encode 'value' as an integer if possible (if integer will fit the
  58  * supported range). If the function sucessful encoded the integer
  59  * then the (up to 5 bytes) encoded representation is written in the
  60  * string pointed by 'enc' and the length is returned. Otherwise
  61  * 0 is returned. */
  62 int rdbEncodeInteger(long long value, unsigned char *enc) {
  63     /* Finally check if it fits in our ranges */
  64     if (value >= -(1<<7) && value <= (1<<7)-1) {
  65         enc[0] = (REDIS_RDB_ENCVAL<<6)|REDIS_RDB_ENC_INT8;
  66         enc[1] = value&0xFF;
  67         return 2;
  68     } else if (value >= -(1<<15) && value <= (1<<15)-1) {
  69         enc[0] = (REDIS_RDB_ENCVAL<<6)|REDIS_RDB_ENC_INT16;
  70         enc[1] = value&0xFF;
  71         enc[2] = (value>>8)&0xFF;
  72         return 3;
  73     } else if (value >= -((long long)1<<31) && value <= ((long long)1<<31)-1) {
  74         enc[0] = (REDIS_RDB_ENCVAL<<6)|REDIS_RDB_ENC_INT32;
  75         enc[1] = value&0xFF;
  76         enc[2] = (value>>8)&0xFF;
  77         enc[3] = (value>>16)&0xFF;
  78         enc[4] = (value>>24)&0xFF;
  79         return 5;
  80     } else {
  81         return 0;
  82     }
  83 }
  84
  85 /* String objects in the form "2391" "-100" without any space and with a
  86  * range of values that can fit in an 8, 16 or 32 bit signed value can be
  87  * encoded as integers to save space */
  88 int rdbTryIntegerEncoding(char *s, size_t len, unsigned char *enc) {
  89     long long value;
  90     char *endptr, buf[32];
  91
  92     /* Check if it's possible to encode this value as a number */
  93     value = strtoll(s, &endptr, 10);
  94     if (endptr[0] != '\0') return 0;
  95     ll2string(buf,32,value);
  96
  97     /* If the number converted back into a string is not identical
  98      * then it's not possible to encode the string as integer */
  99     if (strlen(buf) != len || memcmp(buf,s,len)) return 0;
 100
 101     return rdbEncodeInteger(value,enc);
 102 }
 103
 104 int rdbSaveLzfStringObject(FILE *fp, unsigned char *s, size_t len) {
 105     size_t comprlen, outlen;
 106     unsigned char byte;
 107     int n, nwritten = 0;
 108     void *out;
 109
 110     /* We require at least four bytes compression for this to be worth it */
 111     if (len <= 4) return 0;
 112     outlen = len-4;
 113     if ((out = zmalloc(outlen+1)) == NULL) return 0;
 114     comprlen = lzf_compress(s, len, out, outlen);
 115     if (comprlen == 0) {
 116         zfree(out);
 117         return 0;
 118     }
 119     /* Data compressed! Let's save it on disk */
 120     byte = (REDIS_RDB_ENCVAL<<6)|REDIS_RDB_ENC_LZF;
 121     if ((n = rdbWriteRaw(fp,&byte,1)) == -1) goto writeerr;
 122     nwritten += n;
 123
 124     if ((n = rdbSaveLen(fp,comprlen)) == -1) goto writeerr;
 125     nwritten += n;
 126
 127     if ((n = rdbSaveLen(fp,len)) == -1) goto writeerr;
 128     nwritten += n;
 129
 130     if ((n = rdbWriteRaw(fp,out,comprlen)) == -1) goto writeerr;
 131     nwritten += n;
 132
 133     zfree(out);
 134     return nwritten;
 135
 136 writeerr:
 137     zfree(out);
 138     return -1;
 139 }
 140
 141 /* Save a string objet as [len][data] on disk. If the object is a string
 142  * representation of an integer value we try to safe it in a special form */
 143 int rdbSaveRawString(FILE *fp, unsigned char *s, size_t len) {
 144     int enclen;
 145     int n, nwritten = 0;
 146
 147     /* Try integer encoding */
 148     if (len <= 11) {
 149         unsigned char buf[5];
 150         if ((enclen = rdbTryIntegerEncoding((char*)s,len,buf)) > 0) {
 151             if (rdbWriteRaw(fp,buf,enclen) == -1) return -1;
 152             return enclen;
 153         }
 154     }
 155
 156     /* Try LZF compression - under 20 bytes it's unable to compress even
 157      * aaaaaaaaaaaaaaaaaa so skip it */
 158     if (server.rdbcompression && len > 20) {
 159         n = rdbSaveLzfStringObject(fp,s,len);
 160         if (n == -1) return -1;
 161         if (n > 0) return n;
 162         /* Return value of 0 means data can't be compressed, save the old way */
 163     }
 164
 165     /* Store verbatim */
 166     if ((n = rdbSaveLen(fp,len)) == -1) return -1;
 167     nwritten += n;
 168     if (len > 0) {
 169         if (rdbWriteRaw(fp,s,len) == -1) return -1;
 170         nwritten += len;
 171     }
 172     return nwritten;
 173 }
 174
 175 /* Save a long long value as either an encoded string or a string. */
 176 int rdbSaveLongLongAsStringObject(FILE *fp, long long value) {
 177     unsigned char buf[32];
 178     int n, nwritten = 0;
 179     int enclen = rdbEncodeInteger(value,buf);
 180     if (enclen > 0) {
 181         return rdbWriteRaw(fp,buf,enclen);
 182     } else {
 183         /* Encode as string */
 184         enclen = ll2string((char*)buf,32,value);
 185         redisAssert(enclen < 32);
 186         if ((n = rdbSaveLen(fp,enclen)) == -1) return -1;
 187         nwritten += n;
 188         if ((n = rdbWriteRaw(fp,buf,enclen)) == -1) return -1;
 189         nwritten += n;
 190     }
 191     return nwritten;
 192 }
 193
 194 /* Like rdbSaveStringObjectRaw() but handle encoded objects */
 195 int rdbSaveStringObject(FILE *fp, robj *obj) {
 196     /* Avoid to decode the object, then encode it again, if the
 197      * object is alrady integer encoded. */
 198     if (obj->encoding == REDIS_ENCODING_INT) {
 199         return rdbSaveLongLongAsStringObject(fp,(long)obj->ptr);
 200     } else {
 201         redisAssert(obj->encoding == REDIS_ENCODING_RAW);
 202         return rdbSaveRawString(fp,obj->ptr,sdslen(obj->ptr));
 203     }
 204 }
 205
 206 /* Save a double value. Doubles are saved as strings prefixed by an unsigned
 207  * 8 bit integer specifing the length of the representation.
 208  * This 8 bit integer has special values in order to specify the following
 209  * conditions:
 210  * 253: not a number
 211  * 254: + inf
 212  * 255: - inf
 213  */
 214 int rdbSaveDoubleValue(FILE *fp, double val) {
 215     unsigned char buf[128];
 216     int len;
 217
 218     if (isnan(val)) {
 219         buf[0] = 253;
 220         len = 1;
 221     } else if (!isfinite(val)) {
 222         len = 1;
 223         buf[0] = (val < 0) ? 255 : 254;
 224     } else {
 225 #if (DBL_MANT_DIG >= 52) && (LLONG_MAX == 0x7fffffffffffffffLL)
 226         /* Check if the float is in a safe range to be casted into a
 227          * long long. We are assuming that long long is 64 bit here.
 228          * Also we are assuming that there are no implementations around where
 229          * double has precision < 52 bit.
 230          *
 231          * Under this assumptions we test if a double is inside an interval
 232          * where casting to long long is safe. Then using two castings we
 233          * make sure the decimal part is zero. If all this is true we use
 234          * integer printing function that is much faster. */
 235         double min = -4503599627370495; /* (2^52)-1 */
 236         double max = 4503599627370496; /* -(2^52) */
 237         if (val > min && val < max && val == ((double)((long long)val)))
 238             ll2string((char*)buf+1,sizeof(buf),(long long)val);
 239         else
 240 #endif
 241             snprintf((char*)buf+1,sizeof(buf)-1,"%.17g",val);
 242         buf[0] = strlen((char*)buf+1);
 243         len = buf[0]+1;
 244     }
 245     return rdbWriteRaw(fp,buf,len);
 246 }
 247
 248 /* Save a Redis object. */
 249 int rdbSaveObject(FILE *fp, robj *o) {
 250     int n, nwritten = 0;
 251
 252     if (o->type == REDIS_STRING) {
 253         /* Save a string value */
 254         if ((n = rdbSaveStringObject(fp,o)) == -1) return -1;
 255         nwritten += n;
 256     } else if (o->type == REDIS_LIST) {
 257         /* Save a list value */
 258         if (o->encoding == REDIS_ENCODING_ZIPLIST) {
 259             unsigned char *p;
 260             unsigned char *vstr;
 261             unsigned int vlen;
 262             long long vlong;
 263
 264             if ((n = rdbSaveLen(fp,ziplistLen(o->ptr))) == -1) return -1;
 265             nwritten += n;
 266
 267             p = ziplistIndex(o->ptr,0);
 268             while(ziplistGet(p,&vstr,&vlen,&vlong)) {
 269                 if (vstr) {
 270                     if ((n = rdbSaveRawString(fp,vstr,vlen)) == -1)
 271                         return -1;
 272                     nwritten += n;
 273                 } else {
 274                     if ((n = rdbSaveLongLongAsStringObject(fp,vlong)) == -1)
 275                         return -1;
 276                     nwritten += n;
 277                 }
 278                 p = ziplistNext(o->ptr,p);
 279             }
 280         } else if (o->encoding == REDIS_ENCODING_LINKEDLIST) {
 281             list *list = o->ptr;
 282             listIter li;
 283             listNode *ln;
 284
 285             if ((n = rdbSaveLen(fp,listLength(list))) == -1) return -1;
 286             nwritten += n;
 287
 288             listRewind(list,&li);
 289             while((ln = listNext(&li))) {
 290                 robj *eleobj = listNodeValue(ln);
 291                 if ((n = rdbSaveStringObject(fp,eleobj)) == -1) return -1;
 292                 nwritten += n;
 293             }
 294         } else {
 295             redisPanic("Unknown list encoding");
 296         }
 297     } else if (o->type == REDIS_SET) {
 298         /* Save a set value */
 299         if (o->encoding == REDIS_ENCODING_HT) {
 300             dict *set = o->ptr;
 301             dictIterator *di = dictGetIterator(set);
 302             dictEntry *de;
 303
 304             if ((n = rdbSaveLen(fp,dictSize(set))) == -1) return -1;
 305             nwritten += n;
 306
 307             while((de = dictNext(di)) != NULL) {
 308                 robj *eleobj = dictGetEntryKey(de);
 309                 if ((n = rdbSaveStringObject(fp,eleobj)) == -1) return -1;
 310                 nwritten += n;
 311             }
 312             dictReleaseIterator(di);
 313         } else if (o->encoding == REDIS_ENCODING_INTSET) {
 314             intset *is = o->ptr;
 315             int64_t llval;
 316             int i = 0;
 317
 318             if ((n = rdbSaveLen(fp,intsetLen(is))) == -1) return -1;
 319             nwritten += n;
 320
 321             while(intsetGet(is,i++,&llval)) {
 322                 if ((n = rdbSaveLongLongAsStringObject(fp,llval)) == -1) return -1;
 323                 nwritten += n;
 324             }
 325         } else {
 326             redisPanic("Unknown set encoding");
 327         }
 328     } else if (o->type == REDIS_ZSET) {
 329         /* Save a set value */
 330         zset *zs = o->ptr;
 331         dictIterator *di = dictGetIterator(zs->dict);
 332         dictEntry *de;
 333
 334         if ((n = rdbSaveLen(fp,dictSize(zs->dict))) == -1) return -1;
 335         nwritten += n;
 336
 337         while((de = dictNext(di)) != NULL) {
 338             robj *eleobj = dictGetEntryKey(de);
 339             double *score = dictGetEntryVal(de);
 340
 341             if ((n = rdbSaveStringObject(fp,eleobj)) == -1) return -1;
 342             nwritten += n;
 343             if ((n = rdbSaveDoubleValue(fp,*score)) == -1) return -1;
 344             nwritten += n;
 345         }
 346         dictReleaseIterator(di);
 347     } else if (o->type == REDIS_HASH) {
 348         /* Save a hash value */
 349         if (o->encoding == REDIS_ENCODING_ZIPMAP) {
 350             unsigned char *p = zipmapRewind(o->ptr);
 351             unsigned int count = zipmapLen(o->ptr);
 352             unsigned char *key, *val;
 353             unsigned int klen, vlen;
 354
 355             if ((n = rdbSaveLen(fp,count)) == -1) return -1;
 356             nwritten += n;
 357
 358             while((p = zipmapNext(p,&key,&klen,&val,&vlen)) != NULL) {
 359                 if ((n = rdbSaveRawString(fp,key,klen)) == -1) return -1;
 360                 nwritten += n;
 361                 if ((n = rdbSaveRawString(fp,val,vlen)) == -1) return -1;
 362                 nwritten += n;
 363             }
 364         } else {
 365             dictIterator *di = dictGetIterator(o->ptr);
 366             dictEntry *de;
 367
 368             if ((n = rdbSaveLen(fp,dictSize((dict*)o->ptr))) == -1) return -1;
 369             nwritten += n;
 370
 371             while((de = dictNext(di)) != NULL) {
 372                 robj *key = dictGetEntryKey(de);
 373                 robj *val = dictGetEntryVal(de);
 374
 375                 if ((n = rdbSaveStringObject(fp,key)) == -1) return -1;
 376                 nwritten += n;
 377                 if ((n = rdbSaveStringObject(fp,val)) == -1) return -1;
 378                 nwritten += n;
 379             }
 380             dictReleaseIterator(di);
 381         }
 382     } else {
 383         redisPanic("Unknown object type");
 384     }
 385     return nwritten;
 386 }
 387
 388 /* Return the length the object will have on disk if saved with
 389  * the rdbSaveObject() function. Currently we use a trick to get
 390  * this length with very little changes to the code. In the future
 391  * we could switch to a faster solution. */
 392 off_t rdbSavedObjectLen(robj *o, FILE *fp) {
 393     int nwritten;
 394     if (fp == NULL) fp = server.devnull;
 395     rewind(fp);
 396
 397     /* Determining the saved length of an object should never return -1 */
 398     redisAssert((nwritten = rdbSaveObject(fp,o)) != -1);
 399     return nwritten;
 400 }
 401
 402 /* Return the number of pages required to save this object in the swap file */
 403 off_t rdbSavedObjectPages(robj *o, FILE *fp) {
 404     off_t bytes = rdbSavedObjectLen(o,fp);
 405
 406     return (bytes+(server.vm_page_size-1))/server.vm_page_size;
 407 }
 408
 409 /* Save the DB on disk. Return REDIS_ERR on error, REDIS_OK on success */
 410 int rdbSave(char *filename) {
 411     dictIterator *di = NULL;
 412     dictEntry *de;
 413     FILE *fp;
 414     char tmpfile[256];
 415     int j;
 416     time_t now = time(NULL);
 417
 418     /* Wait for I/O therads to terminate, just in case this is a
 419      * foreground-saving, to avoid seeking the swap file descriptor at the
 420      * same time. */
 421     if (server.vm_enabled)
 422         waitEmptyIOJobsQueue();
 423
 424     snprintf(tmpfile,256,"temp-%d.rdb", (int) getpid());
 425     fp = fopen(tmpfile,"w");
 426     if (!fp) {
 427         redisLog(REDIS_WARNING, "Failed saving the DB: %s", strerror(errno));
 428         return REDIS_ERR;
 429     }
 430     if (fwrite("REDIS0001",9,1,fp) == 0) goto werr;
 431     for (j = 0; j < server.dbnum; j++) {
 432         redisDb *db = server.db+j;
 433         dict *d = db->dict;
 434         if (dictSize(d) == 0) continue;
 435         di = dictGetIterator(d);
 436         if (!di) {
 437             fclose(fp);
 438             return REDIS_ERR;
 439         }
 440
 441         /* Write the SELECT DB opcode */
 442         if (rdbSaveType(fp,REDIS_SELECTDB) == -1) goto werr;
 443         if (rdbSaveLen(fp,j) == -1) goto werr;
 444
 445         /* Iterate this DB writing every entry */
 446         while((de = dictNext(di)) != NULL) {
 447             sds keystr = dictGetEntryKey(de);
 448             robj key, *o = dictGetEntryVal(de);
 449             time_t expiretime;
 450
 451             initStaticStringObject(key,keystr);
 452             expiretime = getExpire(db,&key);
 453
 454             /* Save the expire time */
 455             if (expiretime != -1) {
 456                 /* If this key is already expired skip it */
 457                 if (expiretime < now) continue;
 458                 if (rdbSaveType(fp,REDIS_EXPIRETIME) == -1) goto werr;
 459                 if (rdbSaveTime(fp,expiretime) == -1) goto werr;
 460             }
 461             /* Save the key and associated value. This requires special
 462              * handling if the value is swapped out. */
 463             if (!server.vm_enabled || o->storage == REDIS_VM_MEMORY ||
 464                                       o->storage == REDIS_VM_SWAPPING) {
 465                 /* Save type, key, value */
 466                 if (rdbSaveType(fp,o->type) == -1) goto werr;
 467                 if (rdbSaveStringObject(fp,&key) == -1) goto werr;
 468                 if (rdbSaveObject(fp,o) == -1) goto werr;
 469             } else {
 470                 /* REDIS_VM_SWAPPED or REDIS_VM_LOADING */
 471                 robj *po;
 472                 /* Get a preview of the object in memory */
 473                 po = vmPreviewObject(o);
 474                 /* Save type, key, value */
 475                 if (rdbSaveType(fp,po->type) == -1) goto werr;
 476                 if (rdbSaveStringObject(fp,&key) == -1) goto werr;
 477                 if (rdbSaveObject(fp,po) == -1) goto werr;
 478                 /* Remove the loaded object from memory */
 479                 decrRefCount(po);
 480             }
 481         }
 482         dictReleaseIterator(di);
 483     }
 484     /* EOF opcode */
 485     if (rdbSaveType(fp,REDIS_EOF) == -1) goto werr;
 486
 487     /* Make sure data will not remain on the OS's output buffers */
 488     fflush(fp);
 489     fsync(fileno(fp));
 490     fclose(fp);
 491
 492     /* Use RENAME to make sure the DB file is changed atomically only
 493      * if the generate DB file is ok. */
 494     if (rename(tmpfile,filename) == -1) {
 495         redisLog(REDIS_WARNING,"Error moving temp DB file on the final destination: %s", strerror(errno));
 496         unlink(tmpfile);
 497         return REDIS_ERR;
 498     }
 499     redisLog(REDIS_NOTICE,"DB saved on disk");
 500     server.dirty = 0;
 501     server.lastsave = time(NULL);
 502     return REDIS_OK;
 503
 504 werr:
 505     fclose(fp);
 506     unlink(tmpfile);
 507     redisLog(REDIS_WARNING,"Write error saving DB on disk: %s", strerror(errno));
 508     if (di) dictReleaseIterator(di);
 509     return REDIS_ERR;
 510 }
 511
 512 int rdbSaveBackground(char *filename) {
 513     pid_t childpid;
 514
 515     if (server.bgsavechildpid != -1) return REDIS_ERR;
 516     if (server.vm_enabled) waitEmptyIOJobsQueue();
 517     server.dirty_before_bgsave = server.dirty;
 518     if ((childpid = fork()) == 0) {
 519         /* Child */
 520         if (server.vm_enabled) vmReopenSwapFile();
 521         if (server.ipfd > 0) close(server.ipfd);
 522         if (server.sofd > 0) close(server.sofd);
 523         if (rdbSave(filename) == REDIS_OK) {
 524             _exit(0);
 525         } else {
 526             _exit(1);
 527         }
 528     } else {
 529         /* Parent */
 530         if (childpid == -1) {
 531             redisLog(REDIS_WARNING,"Can't save in background: fork: %s",
 532                 strerror(errno));
 533             return REDIS_ERR;
 534         }
 535         redisLog(REDIS_NOTICE,"Background saving started by pid %d",childpid);
 536         server.bgsavechildpid = childpid;
 537         updateDictResizePolicy();
 538         return REDIS_OK;
 539     }
 540     return REDIS_OK; /* unreached */
 541 }
 542
 543 void rdbRemoveTempFile(pid_t childpid) {
 544     char tmpfile[256];
 545
 546     snprintf(tmpfile,256,"temp-%d.rdb", (int) childpid);
 547     unlink(tmpfile);
 548 }
 549
 550 int rdbLoadType(FILE *fp) {
 551     unsigned char type;
 552     if (fread(&type,1,1,fp) == 0) return -1;
 553     return type;
 554 }
 555
 556 time_t rdbLoadTime(FILE *fp) {
 557     int32_t t32;
 558     if (fread(&t32,4,1,fp) == 0) return -1;
 559     return (time_t) t32;
 560 }
 561
 562 /* Load an encoded length from the DB, see the REDIS_RDB_* defines on the top
 563  * of this file for a description of how this are stored on disk.
 564  *
 565  * isencoded is set to 1 if the readed length is not actually a length but
 566  * an "encoding type", check the above comments for more info */
 567 uint32_t rdbLoadLen(FILE *fp, int *isencoded) {
 568     unsigned char buf[2];
 569     uint32_t len;
 570     int type;
 571
 572     if (isencoded) *isencoded = 0;
 573     if (fread(buf,1,1,fp) == 0) return REDIS_RDB_LENERR;
 574     type = (buf[0]&0xC0)>>6;
 575     if (type == REDIS_RDB_6BITLEN) {
 576         /* Read a 6 bit len */
 577         return buf[0]&0x3F;
 578     } else if (type == REDIS_RDB_ENCVAL) {
 579         /* Read a 6 bit len encoding type */
 580         if (isencoded) *isencoded = 1;
 581         return buf[0]&0x3F;
 582     } else if (type == REDIS_RDB_14BITLEN) {
 583         /* Read a 14 bit len */
 584         if (fread(buf+1,1,1,fp) == 0) return REDIS_RDB_LENERR;
 585         return ((buf[0]&0x3F)<<8)|buf[1];
 586     } else {
 587         /* Read a 32 bit len */
 588         if (fread(&len,4,1,fp) == 0) return REDIS_RDB_LENERR;
 589         return ntohl(len);
 590     }
 591 }
 592
 593 /* Load an integer-encoded object from file 'fp', with the specified
 594  * encoding type 'enctype'. If encode is true the function may return
 595  * an integer-encoded object as reply, otherwise the returned object
 596  * will always be encoded as a raw string. */
 597 robj *rdbLoadIntegerObject(FILE *fp, int enctype, int encode) {
 598     unsigned char enc[4];
 599     long long val;
 600
 601     if (enctype == REDIS_RDB_ENC_INT8) {
 602         if (fread(enc,1,1,fp) == 0) return NULL;
 603         val = (signed char)enc[0];
 604     } else if (enctype == REDIS_RDB_ENC_INT16) {
 605         uint16_t v;
 606         if (fread(enc,2,1,fp) == 0) return NULL;
 607         v = enc[0]|(enc[1]<<8);
 608         val = (int16_t)v;
 609     } else if (enctype == REDIS_RDB_ENC_INT32) {
 610         uint32_t v;
 611         if (fread(enc,4,1,fp) == 0) return NULL;
 612         v = enc[0]|(enc[1]<<8)|(enc[2]<<16)|(enc[3]<<24);
 613         val = (int32_t)v;
 614     } else {
 615         val = 0; /* anti-warning */
 616         redisPanic("Unknown RDB integer encoding type");
 617     }
 618     if (encode)
 619         return createStringObjectFromLongLong(val);
 620     else
 621         return createObject(REDIS_STRING,sdsfromlonglong(val));
 622 }
 623
 624 robj *rdbLoadLzfStringObject(FILE*fp) {
 625     unsigned int len, clen;
 626     unsigned char *c = NULL;
 627     sds val = NULL;
 628
 629     if ((clen = rdbLoadLen(fp,NULL)) == REDIS_RDB_LENERR) return NULL;
 630     if ((len = rdbLoadLen(fp,NULL)) == REDIS_RDB_LENERR) return NULL;
 631     if ((c = zmalloc(clen)) == NULL) goto err;
 632     if ((val = sdsnewlen(NULL,len)) == NULL) goto err;
 633     if (fread(c,clen,1,fp) == 0) goto err;
 634     if (lzf_decompress(c,clen,val,len) == 0) goto err;
 635     zfree(c);
 636     return createObject(REDIS_STRING,val);
 637 err:
 638     zfree(c);
 639     sdsfree(val);
 640     return NULL;
 641 }
 642
 643 robj *rdbGenericLoadStringObject(FILE*fp, int encode) {
 644     int isencoded;
 645     uint32_t len;
 646     sds val;
 647
 648     len = rdbLoadLen(fp,&isencoded);
 649     if (isencoded) {
 650         switch(len) {
 651         case REDIS_RDB_ENC_INT8:
 652         case REDIS_RDB_ENC_INT16:
 653         case REDIS_RDB_ENC_INT32:
 654             return rdbLoadIntegerObject(fp,len,encode);
 655         case REDIS_RDB_ENC_LZF:
 656             return rdbLoadLzfStringObject(fp);
 657         default:
 658             redisPanic("Unknown RDB encoding type");
 659         }
 660     }
 661
 662     if (len == REDIS_RDB_LENERR) return NULL;
 663     val = sdsnewlen(NULL,len);
 664     if (len && fread(val,len,1,fp) == 0) {
 665         sdsfree(val);
 666         return NULL;
 667     }
 668     return createObject(REDIS_STRING,val);
 669 }
 670
 671 robj *rdbLoadStringObject(FILE *fp) {
 672     return rdbGenericLoadStringObject(fp,0);
 673 }
 674
 675 robj *rdbLoadEncodedStringObject(FILE *fp) {
 676     return rdbGenericLoadStringObject(fp,1);
 677 }
 678
 679 /* For information about double serialization check rdbSaveDoubleValue() */
 680 int rdbLoadDoubleValue(FILE *fp, double *val) {
 681     char buf[128];
 682     unsigned char len;
 683
 684     if (fread(&len,1,1,fp) == 0) return -1;
 685     switch(len) {
 686     case 255: *val = R_NegInf; return 0;
 687     case 254: *val = R_PosInf; return 0;
 688     case 253: *val = R_Nan; return 0;
 689     default:
 690         if (fread(buf,len,1,fp) == 0) return -1;
 691         buf[len] = '\0';
 692         sscanf(buf, "%lg", val);
 693         return 0;
 694     }
 695 }
 696
 697 /* Load a Redis object of the specified type from the specified file.
 698  * On success a newly allocated object is returned, otherwise NULL. */
 699 robj *rdbLoadObject(int type, FILE *fp) {
 700     robj *o, *ele, *dec;
 701     size_t len;
 702     unsigned int i;
 703
 704     redisLog(REDIS_DEBUG,"LOADING OBJECT %d (at %d)\n",type,ftell(fp));
 705     if (type == REDIS_STRING) {
 706         /* Read string value */
 707         if ((o = rdbLoadEncodedStringObject(fp)) == NULL) return NULL;
 708         o = tryObjectEncoding(o);
 709     } else if (type == REDIS_LIST) {
 710         /* Read list value */
 711         if ((len = rdbLoadLen(fp,NULL)) == REDIS_RDB_LENERR) return NULL;
 712
 713         /* Use a real list when there are too many entries */
 714         if (len > server.list_max_ziplist_entries) {
 715             o = createListObject();
 716         } else {
 717             o = createZiplistObject();
 718         }
 719
 720         /* Load every single element of the list */
 721         while(len--) {
 722             if ((ele = rdbLoadEncodedStringObject(fp)) == NULL) return NULL;
 723
 724             /* If we are using a ziplist and the value is too big, convert
 725              * the object to a real list. */
 726             if (o->encoding == REDIS_ENCODING_ZIPLIST &&
 727                 ele->encoding == REDIS_ENCODING_RAW &&
 728                 sdslen(ele->ptr) > server.list_max_ziplist_value)
 729                     listTypeConvert(o,REDIS_ENCODING_LINKEDLIST);
 730
 731             if (o->encoding == REDIS_ENCODING_ZIPLIST) {
 732                 dec = getDecodedObject(ele);
 733                 o->ptr = ziplistPush(o->ptr,dec->ptr,sdslen(dec->ptr),REDIS_TAIL);
 734                 decrRefCount(dec);
 735                 decrRefCount(ele);
 736             } else {
 737                 ele = tryObjectEncoding(ele);
 738                 listAddNodeTail(o->ptr,ele);
 739             }
 740         }
 741     } else if (type == REDIS_SET) {
 742         /* Read list/set value */
 743         if ((len = rdbLoadLen(fp,NULL)) == REDIS_RDB_LENERR) return NULL;
 744
 745         /* Use a regular set when there are too many entries. */
 746         if (len > server.set_max_intset_entries) {
 747             o = createSetObject();
 748             /* It's faster to expand the dict to the right size asap in order
 749              * to avoid rehashing */
 750             if (len > DICT_HT_INITIAL_SIZE)
 751                 dictExpand(o->ptr,len);
 752         } else {
 753             o = createIntsetObject();
 754         }
 755
 756         /* Load every single element of the list/set */
 757         for (i = 0; i < len; i++) {
 758             long long llval;
 759             if ((ele = rdbLoadEncodedStringObject(fp)) == NULL) return NULL;
 760             ele = tryObjectEncoding(ele);
 761
 762             if (o->encoding == REDIS_ENCODING_INTSET) {
 763                 /* Fetch integer value from element */
 764                 if (isObjectRepresentableAsLongLong(ele,&llval) == REDIS_OK) {
 765                     o->ptr = intsetAdd(o->ptr,llval,NULL);
 766                 } else {
 767                     setTypeConvert(o,REDIS_ENCODING_HT);
 768                     dictExpand(o->ptr,len);
 769                 }
 770             }
 771
 772             /* This will also be called when the set was just converted
 773              * to regular hashtable encoded set */
 774             if (o->encoding == REDIS_ENCODING_HT) {
 775                 dictAdd((dict*)o->ptr,ele,NULL);
 776             } else {
 777                 decrRefCount(ele);
 778             }
 779         }
 780     } else if (type == REDIS_ZSET) {
 781         /* Read list/set value */
 782         size_t zsetlen;
 783         zset *zs;
 784
 785         if ((zsetlen = rdbLoadLen(fp,NULL)) == REDIS_RDB_LENERR) return NULL;
 786         o = createZsetObject();
 787         zs = o->ptr;
 788         /* Load every single element of the list/set */
 789         while(zsetlen--) {
 790             robj *ele;
 791             double score;
 792             zskiplistNode *znode;
 793
 794             if ((ele = rdbLoadEncodedStringObject(fp)) == NULL) return NULL;
 795             ele = tryObjectEncoding(ele);
 796             if (rdbLoadDoubleValue(fp,&score) == -1) return NULL;
 797             znode = zslInsert(zs->zsl,score,ele);
 798             dictAdd(zs->dict,ele,&znode->score);
 799             incrRefCount(ele); /* added to skiplist */
 800         }
 801     } else if (type == REDIS_HASH) {
 802         size_t hashlen;
 803
 804         if ((hashlen = rdbLoadLen(fp,NULL)) == REDIS_RDB_LENERR) return NULL;
 805         o = createHashObject();
 806         /* Too many entries? Use an hash table. */
 807         if (hashlen > server.hash_max_zipmap_entries)
 808             convertToRealHash(o);
 809         /* Load every key/value, then set it into the zipmap or hash
 810          * table, as needed. */
 811         while(hashlen--) {
 812             robj *key, *val;
 813
 814             if ((key = rdbLoadEncodedStringObject(fp)) == NULL) return NULL;
 815             if ((val = rdbLoadEncodedStringObject(fp)) == NULL) return NULL;
 816             /* If we are using a zipmap and there are too big values
 817              * the object is converted to real hash table encoding. */
 818             if (o->encoding != REDIS_ENCODING_HT &&
 819                ((key->encoding == REDIS_ENCODING_RAW &&
 820                 sdslen(key->ptr) > server.hash_max_zipmap_value) ||
 821                 (val->encoding == REDIS_ENCODING_RAW &&
 822                 sdslen(val->ptr) > server.hash_max_zipmap_value)))
 823             {
 824                     convertToRealHash(o);
 825             }
 826
 827             if (o->encoding == REDIS_ENCODING_ZIPMAP) {
 828                 unsigned char *zm = o->ptr;
 829                 robj *deckey, *decval;
 830
 831                 /* We need raw string objects to add them to the zipmap */
 832                 deckey = getDecodedObject(key);
 833                 decval = getDecodedObject(val);
 834                 zm = zipmapSet(zm,deckey->ptr,sdslen(deckey->ptr),
 835                                   decval->ptr,sdslen(decval->ptr),NULL);
 836                 o->ptr = zm;
 837                 decrRefCount(deckey);
 838                 decrRefCount(decval);
 839                 decrRefCount(key);
 840                 decrRefCount(val);
 841             } else {
 842                 key = tryObjectEncoding(key);
 843                 val = tryObjectEncoding(val);
 844                 dictAdd((dict*)o->ptr,key,val);
 845             }
 846         }
 847     } else {
 848         redisPanic("Unknown object type");
 849     }
 850     return o;
 851 }
 852
 853 /* Mark that we are loading in the global state and setup the fields
 854  * needed to provide loading stats. */
 855 void startLoading(FILE *fp) {
 856     struct stat sb;
 857
 858     /* Load the DB */
 859     server.loading = 1;
 860     server.loading_start_time = time(NULL);
 861     if (fstat(fileno(fp), &sb) == -1) {
 862         server.loading_total_bytes = 1; /* just to avoid division by zero */
 863     } else {
 864         server.loading_total_bytes = sb.st_size;
 865     }
 866 }
 867
 868 /* Refresh the loading progress info */
 869 void loadingProgress(off_t pos) {
 870     server.loading_loaded_bytes = pos;
 871 }
 872
 873 /* Loading finished */
 874 void stopLoading(void) {
 875     server.loading = 0;
 876 }
 877
 878 int rdbLoad(char *filename) {
 879     FILE *fp;
 880     uint32_t dbid;
 881     int type, retval, rdbver;
 882     int swap_all_values = 0;
 883     redisDb *db = server.db+0;
 884     char buf[1024];
 885     time_t expiretime, now = time(NULL);
 886     long loops = 0;
 887
 888     fp = fopen(filename,"r");
 889     if (!fp) return REDIS_ERR;
 890     if (fread(buf,9,1,fp) == 0) goto eoferr;
 891     buf[9] = '\0';
 892     if (memcmp(buf,"REDIS",5) != 0) {
 893         fclose(fp);
 894         redisLog(REDIS_WARNING,"Wrong signature trying to load DB from file");
 895         return REDIS_ERR;
 896     }
 897     rdbver = atoi(buf+5);
 898     if (rdbver != 1) {
 899         fclose(fp);
 900         redisLog(REDIS_WARNING,"Can't handle RDB format version %d",rdbver);
 901         return REDIS_ERR;
 902     }
 903
 904     startLoading(fp);
 905     while(1) {
 906         robj *key, *val;
 907         int force_swapout;
 908
 909         expiretime = -1;
 910
 911         /* Serve the clients from time to time */
 912         if (!(loops++ % 1000)) {
 913             loadingProgress(ftello(fp));
 914             aeProcessEvents(server.el, AE_FILE_EVENTS|AE_DONT_WAIT);
 915         }
 916
 917         /* Read type. */
 918         if ((type = rdbLoadType(fp)) == -1) goto eoferr;
 919         if (type == REDIS_EXPIRETIME) {
 920             if ((expiretime = rdbLoadTime(fp)) == -1) goto eoferr;
 921             /* We read the time so we need to read the object type again */
 922             if ((type = rdbLoadType(fp)) == -1) goto eoferr;
 923         }
 924         if (type == REDIS_EOF) break;
 925         /* Handle SELECT DB opcode as a special case */
 926         if (type == REDIS_SELECTDB) {
 927             if ((dbid = rdbLoadLen(fp,NULL)) == REDIS_RDB_LENERR)
 928                 goto eoferr;
 929             if (dbid >= (unsigned)server.dbnum) {
 930                 redisLog(REDIS_WARNING,"FATAL: Data file was created with a Redis server configured to handle more than %d databases. Exiting\n", server.dbnum);
 931                 exit(1);
 932             }
 933             db = server.db+dbid;
 934             continue;
 935         }
 936         /* Read key */
 937         if ((key = rdbLoadStringObject(fp)) == NULL) goto eoferr;
 938         /* Read value */
 939         if ((val = rdbLoadObject(type,fp)) == NULL) goto eoferr;
 940         /* Check if the key already expired */
 941         if (expiretime != -1 && expiretime < now) {
 942             decrRefCount(key);
 943             decrRefCount(val);
 944             continue;
 945         }
 946         /* Add the new object in the hash table */
 947         retval = dbAdd(db,key,val);
 948         if (retval == REDIS_ERR) {
 949             redisLog(REDIS_WARNING,"Loading DB, duplicated key (%s) found! Unrecoverable error, exiting now.", key->ptr);
 950             exit(1);
 951         }
 952         /* Set the expire time if needed */
 953         if (expiretime != -1) setExpire(db,key,expiretime);
 954
 955         /* Handle swapping while loading big datasets when VM is on */
 956
 957         /* If we detecter we are hopeless about fitting something in memory
 958          * we just swap every new key on disk. Directly...
 959          * Note that's important to check for this condition before resorting
 960          * to random sampling, otherwise we may try to swap already
 961          * swapped keys. */
 962         if (swap_all_values) {
 963             dictEntry *de = dictFind(db->dict,key->ptr);
 964
 965             /* de may be NULL since the key already expired */
 966             if (de) {
 967                 vmpointer *vp;
 968                 val = dictGetEntryVal(de);
 969
 970                 if (val->refcount == 1 &&
 971                     (vp = vmSwapObjectBlocking(val)) != NULL)
 972                     dictGetEntryVal(de) = vp;
 973             }
 974             decrRefCount(key);
 975             continue;
 976         }
 977         decrRefCount(key);
 978
 979         /* Flush data on disk once 32 MB of additional RAM are used... */
 980         force_swapout = 0;
 981         if ((zmalloc_used_memory() - server.vm_max_memory) > 1024*1024*32)
 982             force_swapout = 1;
 983
 984         /* If we have still some hope of having some value fitting memory
 985          * then we try random sampling. */
 986         if (!swap_all_values && server.vm_enabled && force_swapout) {
 987             while (zmalloc_used_memory() > server.vm_max_memory) {
 988                 if (vmSwapOneObjectBlocking() == REDIS_ERR) break;
 989             }
 990             if (zmalloc_used_memory() > server.vm_max_memory)
 991                 swap_all_values = 1; /* We are already using too much mem */
 992         }
 993     }
 994     fclose(fp);
 995     stopLoading();
 996     return REDIS_OK;
 997
 998 eoferr: /* unexpected end of file is handled here with a fatal exit */
 999     redisLog(REDIS_WARNING,"Short read or OOM loading DB. Unrecoverable error, aborting now.");
1000     exit(1);
1001     return REDIS_ERR; /* Just to avoid warning */
1002 }
1003
1004 /* A background saving child (BGSAVE) terminated its work. Handle this. */
1005 void backgroundSaveDoneHandler(int statloc) {
1006     int exitcode = WEXITSTATUS(statloc);
1007     int bysignal = WIFSIGNALED(statloc);
1008
1009     if (!bysignal && exitcode == 0) {
1010         redisLog(REDIS_NOTICE,
1011             "Background saving terminated with success");
1012         server.dirty = server.dirty - server.dirty_before_bgsave;
1013         server.lastsave = time(NULL);
1014     } else if (!bysignal && exitcode != 0) {
1015         redisLog(REDIS_WARNING, "Background saving error");
1016     } else {
1017         redisLog(REDIS_WARNING,
1018             "Background saving terminated by signal %d", WTERMSIG(statloc));
1019         rdbRemoveTempFile(server.bgsavechildpid);
1020     }
1021     server.bgsavechildpid = -1;
1022     /* Possibly there are slaves waiting for a BGSAVE in order to be served
1023      * (the first stage of SYNC is a bulk transfer of dump.rdb) */
1024     updateSlavesWaitingBgsave(exitcode == 0 ? REDIS_OK : REDIS_ERR);
1025 }